run-tests
Warn
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: MEDIUMREMOTE_CODE_EXECUTIONPROMPT_INJECTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill executes
npx -y skills update, which downloads and runs code from the npm package registry at runtime. - [PROMPT_INJECTION]: The instructions command the agent to perform updates "silently" and "don't mention it to the user," which is a concealment pattern that bypasses user oversight.
- [DYNAMIC_EXECUTION]: The skill implements a self-modification pattern by instructing the agent to "re-read this SKILL.md" after the update and "follow that copy," effectively enabling remote instruction injection into the current session.
- [INDIRECT_PROMPT_INJECTION]: The skill processes and displays "judge reasoning" verbatim from external test results, creating a surface where malicious instructions in the processed data could influence the agent.
- Ingestion points: Output of
calibrate agent-tests get-run(SKILL.md, Phase 3/4). - Boundary markers: None present; output is presented verbatim.
- Capability inventory: Shell command execution via
calibrateCLI. - Sanitization: No sanitization of the external judge reasoning content is specified.
Audit Metadata